Upload
trinhkien
View
218
Download
4
Embed Size (px)
Citation preview
Los Angeles, CA ▪ San Francisco, CA ▪ Tallahassee, FL ▪ Washington, DC © 2011, ERSGroup
Regression Analysis Applications in Litigation
By
Robert Mills
Director
Micronomics, Inc.
Dubravka Tosic, Ph.D.
Principal
ERS Group
Practising Law Institute
Pocket MBA: Finance for Lawyers
Summer 2011
-1-
Regression Analysis Applications in Litigation
Robert Mills*
Dubravka Tosic, Ph.D.**
March 2011
I. Introduction to Regression Analysis
Regression analysis is a statistical tool used to examine
relationships among variables. It provides a method for
quantifying the impact of changes in one or more explanatory
variables (known as independent variables) on a variable of
interest (known as the dependent variable). Regression analysis is
widely used in the field of econometrics, which is concerned with
the application of statistical and mathematical methods to the
analysis of economic data.1 Useful applications also are found in
finance, sociology, biology, psychology, pharmacology, and
engineering, among other fields of study. In this paper, we provide
an introduction to regression analysis and discuss a number of
applications in the litigation context.
Regression analysis begins with a hypothesis. Suppose, for
example, that we are interested in understanding factors that
impact attendance at a sporting event. We might hypothesize that
historical performance of the home team influences attendance.
* Robert Mills is a Director at Micronomics, Inc., an economic research
and consulting firm in Los Angeles, California. Micronomics is a
subsidiary of ERS Group, a national economic and statistical consulting
firm. **
Dubravka Tosic, Ph.D. is a Principal at ERS Group, and based in New
York/New Jersey. 1 Additional information can be found in an econometrics textbook such
as James H. Stock and Mark W. Watson, Introduction to Econometrics,
3rd ed. (Upper Saddle River: Prentice-Hall, 2010); William H. Greene,
Econometric Analysis, 7th ed. (Upper Saddle River: Prentice-Hall, 2011);
or Peter Kennedy, A Guide to Econometrics, 5th ed. (Cambridge: The
MIT Press, 2003).
-2-
We might further believe that the relationship between historical
performance and attendance is positive; that is, improvements in
performance of the home team lead to greater attendance and
declines in performance of the home team lead to lower
attendance. Assuming historical attendance and home team
performance data are available, we can estimate the following
model:
where
= attendance at game i (the dependent variable);
= home team performance as of game i measured by the
win-loss record expressed as a percentage (the
independent variable);
= constant amount (interpreted as attendance given a
win-loss record of zero percent);
= the effect in attendance of each additional percentage
in the home team win-loss record; and
= a “disturbance” term reflecting other unmeasured
factors that influence attendance.
Data for A and P are plotted in the following figure. The
coefficients and are not known. Regression analysis produces
estimates for these coefficients, which customarily are denoted
with a “hat” superscript (e.g., and ). The disturbance term, ,
also is unknown.
-3-
Graphically, estimation of the coefficients and is tantamount
to fitting a line to the attendance and home team win-loss record
data, where is the point at which the line intersects the vertical
axis and is the slope of the line. The following figure depicts
such a line.
This line appears to fit the data. Without an objective criterion,
however, there is no guarantee that this line provides the best fit.
Regression analysis provides a criterion. With regression analysis,
Att
end
ance
Home Team Win-Loss Record (%)
Att
endan
ce
Home Team Win-Loss Record (%)
-4-
the intercept and slope of the line (i.e., and ) typically are
estimated by minimizing the sum of squared errors (“SSE”).
First, an estimated error for each observation is measured as the
vertical distance between the observed value of the variable and
the estimated line. SSE is calculated by squaring this estimated
error for each observation and summing across all observations.
Estimates of the coefficients are chosen to minimize SSE. This is
called the method of ordinary least squares. In practice, this
estimation is carried out using regression software. With ordinary
least squares the best fitted line for the data is estimated.
Common knowledge suggests that attendance at sporting events
increases with improvements in home team performance. In other
words, we expect a positive coefficient for home team win-loss
record ( indicating that attendance increases as performance
improves and attendance decreases as performance declines, other
things equal.
Estimating our model produces the following results.
Regression Output
R2 = 0.70
Coefficient
Standard
Error
t-Statistic
Intercept ( 25,419 4,913 5.17
Win-Loss Record ( ) 501 90 5.55
The estimated coefficient for the home team win-loss record is
501, which is interpreted as the estimated number of additional
attendees for every one percent improvement in the home team
win-loss record. This estimate is consistent with our expectation
that the coefficient is positive. The intercept term is interpreted as
the estimated number of attendees given a home team record of
zero wins. Using these coefficient estimates, attendance can be
predicted for any given home team win-loss record. For example,
-5-
if the win-loss record is 50% as of game i, estimated attendance at
game i is
50,469 = 25,419 + (501 * 50).
The model suggests attendance would increase to 62,994 in the
event that the home team win-loss record improved to 75%:
62,994 = 25,419 + (501 * 75).
The results of the regression analysis appear to confirm our a
priori belief that attendance increases with improvements in home
team performance. Using the t-statistic reported above, we can
formally test the hypothesis that performance does not impact
attendance. Operationally, this test involves comparing the
reported t-statistic for the coefficient of interest to the critical value
obtained from the t distribution. Courts have frequently adopted
the concept of statistical “significance” when assessing the
importance of a variable. Assuming a large sample size, the critical
value is 1.96 (or approximately two standard deviations) at the five
percent level of significance. Since the reported t-statistic of 5.55
exceeds the critical value of 1.96, we can reject the hypothesis that
performance does not impact attendance at a five percent level of
statistical significance.
Another useful statistic frequently reported with regression results
is the coefficient of determination, or R-squared (R2). R
2 reflects
the proportion of total variation in the dependent variable
explained by variation in the independent variable or variables. In
other words, it provides a measure of the “explanatory power” of a
model.
The value of R2 ranges from 0 to 1, with a value of 0 meaning that
none of the variation in the dependent variable is explained by
variation in the independent variables and a value of 1 signifying
that all of the variation in the dependent variable is explained by
variation in the independent variables. Roughly speaking, a high
value of R2 often is associated with a good fit of the regression line
whereas a low value of R2 is associated with a poor fit. This does
not mean, however, that the relative strength of two competing
-6-
models can be assessed by their respective R2 values alone. Indeed,
introducing additional independent variables into a model will tend
to increase the value of R2 even where those variables have no
hypothesized relationship with the dependent variable.
Specification of the regression model should be founded in theory.
Explanatory variables should not be included in a model without a
theoretical basis for inclusion. Similarly, explanatory variables that
theory suggests are relevant should not be excluded without a
sound basis for doing so. The exclusion of relevant explanatory
variables from a model without basis is particularly problematic
because it can lead to omitted variable bias.
Omitted variable bias arises when one or more variables that
should be included in a model are excluded from the estimated
model. In such cases, coefficient estimates for the included
variables can be biased and the results of hypothesis tests rendered
unreliable.2
Turning back to the sporting event attendance example, the only
explanatory variable we have considered is home team
performance as measured by the home team win-loss record.
Clearly, this is an overly simplistic view of the determinants of
attendance. Attendance likely is affected by a variety of factors in
addition to home team performance. Economic theory suggests, for
example, that ticket sales depend in part on ticket prices. The win-
loss record of the visiting team, the number of games left in the
season, the day of the week on which the event occurs, and game
day rainfall (particularly for outdoor events) might also be
relevant. Each of these variables is subject to measurement and
could be included in the model. The problem of omitted variable
bias is more troublesome when the omitted variable is not readily
subject to measurement and therefore cannot be included.
The problem of including irrelevant variables typically is less
serious than the problem of omitting relevant variables because the
2 Coefficient estimates for included variables will remain unbiased in the
event that the omitted variable is uncorrelated with all of the included
variables.
-7-
inclusion of irrelevant variables does not serve to bias estimates of
the coefficients for relevant variables. This is not to say that the
practice of including irrelevant variables in a model is without
cost. The efficiency of the estimator is affected by including
irrelevant variables, which can be problematic particularly when
working with small sample sizes.
While identifying relevant explanatory variables is an essential
aspect of model specification, the choice of functional form also is
important. Thus far we have assumed that the relationship between
the dependent variable and the independent variables is linear.
Depending upon the application, theory may suggest that a
nonlinear functional form is more appropriate. The left and right
panels of the following figure illustrate nonlinear functional forms
commonly used in practice. The left panel depicts a semi-log
model and the right panel depicts a polynomial model.
Many nonlinear functional forms, including those shown above,
can be estimated using standard linear regression techniques
because they are linear in the coefficients. Nonlinear functional
forms that are not subject to linear transformations require more
sophisticated estimation techniques.
There are a number of assumptions that underlie the standard
linear regression model. It is important to recognize situations
-8-
where these assumptions are violated so that alternative methods
can be employed to produce sound results. It is beyond the scope
of this paper to provide a comprehensive overview of all of the
problems that can arise. Instead, we will focus on two common
problems that are related to the disturbance term.
It typically is assumed that the disturbance term is composed of
small, individually unimportant effects that are independently
distributed from a normal population with an expected value of
zero and constant variance.
Violations of the constant variance assumption are not uncommon
in practice. When working with data from a cross section of firms
in an industry, for example, a systematic difference between the
disturbances for large and small firms may exist indicating that
variance of the disturbance term is not constant. Disturbances are
said to be heteroscedastic when they have different variances.
Violations of the independence assumption sometimes arise when
working with time series data because the disturbance associated
with observations for one period may carry over into future
periods. Disturbances are said to be serially correlated when the
disturbance terms for different observations are correlated.
In the presence of heteroscedasticity or serial correlation, the
method of ordinary least squares produces coefficient estimates
that are unbiased but not efficient. The loss of efficiency is
undesirable because it can affect the results of hypothesis tests.
Fortunately, procedures for identifying and correcting problems
associated with heteroscedasticity and serial correlation are readily
available.
II. Examples of Practical Applications of Regression Analysis
The discussion thus far is intended to provide non-practitioners a
brief introduction to regression analysis. We now introduce some
practical applications of regression analysis in the litigation
context. Specifically, we provide an overview of (A) the role of
regression analysis in estimating price elasticity of demand in
antitrust and intellectual property matters, (B) use of regression
-9-
analysis to conduct event studies designed to estimate the impact
of specific events on the value of a firm, (C) the application of
regression analysis to cost estimation in damages studies, and (D)
applications of regression analysis in labor and employment
disputes.
A. Price Elasticity of Demand
Demand refers to the quantity of a good or service consumers
purchase at prevailing prices. Increases in the prevailing price of a
good tend to result in reduced sales volume because some
consumers will choose alternative products or refrain altogether
from making a purchase as price increases. Conversely, decreases
in the prevailing price tend to result in sales volume increases. The
term price elasticity of demand refers to the extent to which sales
volume is affected by price changes.
Own-price elasticity of demand measures the responsiveness of the
quantity of a good demanded to changes in its price. Demand is
said to be elastic if quantity demanded is highly sensitive to
changes in price, and inelastic if price changes have little impact
on quantity demanded. Cross-price elasticity of demand measures
the responsiveness of quantity demanded for one good to changes
in the price of another good.
Own-price elasticity of demand is negative since price increases
lead to decreases in quantity demanded. This elasticity commonly
is reported in terms of absolute value, however, and the negative
sign can be assumed. Cross-price elasticity of demand can be
positive or negative depending upon whether the goods are
substitutes (positive cross-price elasticity) or compliments
(negative cross-price elasticity). Together, own- and cross-price
elasticity summarize anticipated substitution patterns among
consumers faced with changes in price.
The concept of price elasticity of demand has been widely used in
litigation, notably in assessing potential anticompetitive effects of
mergers. Own- and cross-price elasticity are routinely used to
define relevant antitrust markets, assess market power, and
-10-
simulate price increases resulting from mergers before they are
consummated.
Use of price elasticity of demand also has emerged in patent
infringement litigation, particularly in cases where price erosion is
alleged to have occurred. An assessment of price erosion involves
estimating the price that would have prevailed but for the
infringement and then determining the amount of sales the patent
owner would have made at that price. Although a patent owner
may have been able to charge a higher price in the absence of the
infringement, its sales might have been lower depending upon the
price elasticity of demand.
Measures of price elasticity of demand commonly are derived by
estimating one or more demand curves using regression analysis.
Economic theory suggests the quantity of a good demanded
depends upon its price, the price of substitutes and complements,
and income, among other possible factors. In practice, data
limitations may dictate which variables are included in a regression
analysis, but the potential for omitted variable bias also should be
considered when specifying models.
Suppose we have monthly data on the quantity of goods sold (
and ) and corresponding price data ( and ) for two substitute
goods. We also have monthly income data (I) for consumers that
purchase the goods. We can estimate the following linear demand
equations using regression analysis.
We use a linear demand model in this example for simplicity.
Economic theory does not dictate an exact functional relationship
between quantity demanded and the variables that impact demand.
The properties of a specific functional form may lead the
researcher to believe it superior for a given situation, but the
choice is often somewhat arbitrary. If sufficient data are available,
a variety of functional forms might be estimated to assess the
sensitivity of the results to the choice of functional form. This
-11-
practice may lend credibility to the results if they are shown to be
insensitive to the choice of functional form. Results that are
extremely sensitive to functional form may prove difficult to
defend.
Price elasticity for both goods can readily be estimated using the
estimated coefficients from the linear demand model. Own-price
elasticity is equal to the “first partial derivative” of the demand
equation with respect to price times price divided by quantity.
In other words, own-price elasticity of demand is equal to the
coefficient for the price variable multiplied by price which is
divided by quantity. Cross-price elasticity of demand is calculated
as the coefficient for the price of the other good multiplied by the
price of the other good divided by quantity.
Price elasticity estimates can prove useful in the litigation context,
particularly in cases where the interplay between price and
quantity is an issue. In antitrust litigation, for example, elasticity
and cross-price elasticity are often used to delineate relevant
markets. Firms are likely to be grouped in the same market if the
products they produce can be used interchangeably and where the
products exhibit a high cross-price elasticity of demand. In cases
where price allegedly would have been higher (or lower) in the
absence of some conduct, elasticity estimates can be used to show
the impact of that but-for price on quantity demanded.
-12-
B. Event Study Analysis
Event studies measure the impact of specific events on the value of
firms. There are many useful applications for event studies in
litigation settings. For example, event studies are commonly used
to estimate the impact of adverse information on movements in
share prices in matters of alleged securities fraud. They also can
provide insight into damages resulting from events such as product
recalls, the loss of patent protection, credit facility constraints, and
fraud.
The basic premise underlying an event study analysis is that given
rational market participants, security prices will quickly adjust to
reflect the announcement of an event. Roughly speaking, security
price changes are attributable to company-specific information
(such as the announcement of a new product) and industry or
market-wide information (such as new regulation or changes in
interest rates). Event study analysis provides a framework for
isolating the impact of company-specific events on security prices.
The total impact of an event can then be estimated by summing the
company-specific impact across all of shares affected.
The first step in undertaking an event study analysis involves the
identification of the event or events of interest. In the litigation
context, the events of interest often are dictated by allegations in
the complaint. Suppose, for example, that a publicly traded early-
stage pharmaceutical company alleges that clinical trials for a
potential new therapeutic drug were unsuccessful as a result of a
failure on the part of its development partner to design a proper
test protocol. In this example, the event of interest is the public
announcement that the clinical trials were unsuccessful.
After the event of interest has been identified, it is necessary to
determine the period of time over which the impact will be
measured. This is called the event window. In practice, the event
window typically is defined to include at least the day on which
the event was announced and the following business day.
Depending upon the circumstances, the event window may
commence before the event is announced (e.g., if there is reason to
believe that news of the event leaked before the official
-13-
announcement) and end days after the event is announced (e.g., if
there is reason to believe that some market participants did not
immediately learn of the event at the time it was announced). The
event window ideally will be long enough to include any ongoing
adjustment to news of the event in the market, but not so long as to
capture effects of unrelated subsequent events.
A primary objective of event study analysis is to isolate the impact
of the event in question from market-wide and industry-wide
information that also impacts securities prices. The following
model is often used in this context:
where
the security return on day t for the company of
interest;
the market index return on day t;
the intercept coefficient;
the market index coefficient; and
a disturbance term reflecting other factors that
influence the security return for the company of
interest.
Historical stock price data for the company in question are
collected and daily returns are calculated. Market index data also
are collected. This market index may be a widely available index
such as the Standard and Poor’s 500 or a custom index that
includes peers of the company of interest. Returning to our early-
stage pharmaceutical company example, a useful market index
might be constructed to include other publicly traded early-stage
companies involved in clinical trials for potential new therapeutics.
-14-
Regression analysis is employed to obtain estimates for and .
The results of the regression analysis are then used to calculate the
predicted security return, for each day in the event window:
.
The predicted security return is essentially an estimate of the
security return but for the event in question. Predicted security
returns are compared to actual returns to determine the impact of
the event in question. The difference between the actual and
predicted return on any particular day is called the abnormal
return:
abnormal return .
Summing abnormal returns across all days in the event window
yields cumulative abnormal returns:
cumulative abnormal returns
Cumulative abnormal returns (or CAR) speak to the magnitude of
the event in question.
Event studies can also shed light on the materiality of events.
Materiality is addressed using statistical testing. A common
question in event studies is whether or not the hypothesis that the
cumulative abnormal returns are zero can be rejected. Output
obtained from the regression analysis provides the information
necessary to conduct such a test.
Event study analysis has been used in a wide range of
investigations. In the litigation context, it is used to estimate
damages caused by securities fraud and other wrongful conduct.
Event studies have also been used to understand the value created
by mergers and acquisitions, the impact of corporate earnings
restatements, and market reactions to jury verdicts.
-15-
C. Cost Estimation in Damages Studies
In many cases, lost profits damages are calculated as the difference
between profits that would have been generated but for some
alleged conduct, such as a breach of contract, and actual profits
generated given the conduct. Estimating but-for profits requires an
understanding of the costs involved, and in particular those costs
that were not incurred given the alleged conduct but would have
been incurred in the absence of the alleged conduct. These costs
are sometimes referred to as avoided costs. The estimation of
avoided costs often requires an understanding of the distinction
between those cost elements that are fixed and those that are
variable.
Fixed costs do not vary with levels of output. Costs that frequently
are fixed over moderate changes in output include rent, insurance
premiums, business license fees, and salaries for permanent full
time employees.
Variable costs are those that vary directly with the level of output.
Depending upon the nature of the business, variable costs may
include cost of goods sold, shipping charges, royalties, and sales
commissions, among others.
Certain costs cannot be classified as strictly fixed or variable.
These semi-variable costs include a mixture of fixed and variable
components. Common examples of semi-variable costs include
production labor (regular wages are fixed but overtime is variable),
electricity, telephone bills, and postage.
An important consideration when assessing the nature of costs is
that cost elements can be fixed over certain levels of output and
variable over other levels of output. To illustrate this point,
suppose a manufacturer has the capacity to increase production by
ten percent without expanding its plant, but any increase in
production above ten percent would require an expansion. In this
example, the rent associated with the plant is fixed over relatively
small increases in output. Increasing output by more than ten
percent, however, would require an expansion of the plant and the
-16-
payment of additional rent. In other words, rent is a variable cost in
this example over large increases in output.
Discussions with company management and accounting personnel
can be helpful in understanding the fixed or variable nature of
costs. Depending upon the availability of data, regression analysis
may provide additional insight.
Regression analysis provides a means to examine and quantify
relationships among variables. In the case of cost estimation, a
common inquiry is “what is the relationship between changes in
output and the cost of production?” Assuming sufficient data are
available, the following model might be estimated to address this
question:
where
= cost of production during period t;
= production during period t;
= the intercept coefficient;
= the production coefficient; and
= a disturbance term reflecting other factors that
influence the cost of production.3
The coefficient is interpreted as the cost of production when
output falls to zero units. In other words, it provides an indication
of the fixed cost of production. The coefficient is interpreted as
the cost of production for one additional unit of output. That is, it
provides an indication of the variable cost of production. Together,
3 Depending upon the situation, model specification might be more
complicated in practice. Decisions concerning the variables to include,
functional form, and data aggregation are driven by the specific facts and
circumstances of the investigation.
-17-
these coefficients can be used to estimate the total cost of
production for a given level of output.
In our example, the regression results might be used to calculate
profit but for the alleged conduct:
Profit Sales Price .
Subtracting actual profits from but-for profits would yield an
estimate of profits lost as a result of the alleged conduct.
D. Labor and Employment Litigation
Almost all employers face federal nondiscrimination requirements,
and most states also have enacted employment laws specifically
dealing with discrimination. These federal and state laws are
intended to ensure that employers base employment practices (e.g.,
hiring, promotion, termination, discipline, compensation) on
objective and fair measures, such as performance and merit.
Employment discrimination allegations often charge employers
with engaging in discrimination against a member or members of a
protected class (legally protected characteristics include race,
gender, ethnicity, national origin, religion, age, and disability).
These allegations require plaintiffs to demonstrate that a pattern or
practice of discrimination exists. Statistical analysis is commonly
used to analyze such allegations. Various statistical tests can be
performed utilizing human resources, payroll, and other business
data. Regression analysis can also be employed to identify patterns
in data that reflect employment decisions.
Regression analysis may be viewed as a tool that quantifies the
relationship between a decision variable and other independent
factors. For example, suppose a company faces an employment
discrimination matter in which plaintiffs allege that women are
being discriminated against in terms of base pay. The hypothesis
we would want to test with regression analysis is that gender is not
a significant factor in determining the base salary level of
employees. The following multiple regression model could be
estimated:
-18-
where
base salary for employee n;
characteristics of employee n;
gender of employee n;
the intercept coefficient;
the employee characteristics coefficients;
the gender coefficient; and
a disturbance term reflecting other factors that
influence base salaries.
This model is referred to as a multiple regression model since
multiple explanatory variables are considered. In our example, the
dependent variable is base salary and the independent variables are
various characteristics of employees that might influence base
salary and for which data are available. The employer might
contend that the following employee characteristics are important
determinants of base salary, and as such should be included in the
regression model: education, prior experience, tenure, special
skills, department, and geographic region. To test the hypothesis
that base salary for women is not different than the base salary for
men after controlling for all of these factors, the regression model
would also include a variable that reflects the gender of the
employee, which is depicted in our model as G. The constant term,
, is interpreted as the average base salary paid to a man who has a
zero value in each independent variable (e.g., no education, no
prior experience, and no tenure). The coefficients and measure
the influence of the independent variables on base salary.
Estimates of these coefficients are referred to as unbiased estimates
of the influence of the independent variables on the dependent
variable if the variables are independent of each other, no
-19-
important variables have been omitted, base salary is normally
distributed, and other assumptions underlying the method of
ordinary least squares hold.
The difference between average base salary for men and women is
estimated by the coefficient . If this coefficient is statistically
significant (i.e., it has a t-statistic of more than 1.96 assuming a
five percent level of statistical significance), the difference
between the base salary for men and women is said to be
statistically significant after accounting for other factors included
in the regression model. Assuming the regression model controls
for factors influencing pay, this result would prompt us to reject
the hypothesis that gender is not a significant factor in determining
base salary.
Given the widespread availability of computing power and
sophisticated computer software, it is possible to generate a wealth
of information useful for identifying and examining outliers,
testing the robustness of models, and analyzing the sensitivity of
results to assumptions made. For instance, significant outliers are
often examined to further evaluate the quality of the model and
data. Using the base salary example provided above, data
pertaining to employees that are identified as statistically
significant positive or negative outliers (i.e., employees whose
actual base salary is significantly higher or lower than their
predicted base salary), could be reviewed to identify potential
anomalies in the data. This process can provide information that
might be used to further refine the model.
III. Conclusion
Implementing regression analysis requires an appreciation for the
statistical underpinnings of the analysis along with a well-designed
model that is founded in theory. When used properly, regression
analysis is a powerful tool with many practical applications in
litigation. It has been widely accepted by courts as a reliable
estimation framework.
-20-
About the Authors
ROBERT MILLS is an economist and Director at Micronomics,
Inc., an economic research and consulting firm located in Los
Angeles, California. Micronomics, Inc. is a subsidiary of ERS
Group, a national economic and statistical consulting firm.
Mr. Mills has been engaged in economic research and consulting
for the past 15 years. A significant portion of his professional
experience has involved the valuation of intellectual property and
other assets, industrial organization, and the calculation of
economic damages. His experience spans many industries,
including software, semiconductors, health care, medical devices,
pharmaceuticals, entertainment, telecommunications, real estate,
apparel, manufacturing, retail sales, insurance, sporting goods, and
energy, among others.
Mr. Mills has served as an expert witness or consultant in a wide
range of matters, including patent, trademark and copyright
infringement, theft of trade secrets, breach of contract,
interference, conversion, fraud, predatory pricing, attempted
monopolization, and labor disputes. He has testified as an
economic expert in Federal District Court, state courts in multiple
jurisdictions, and at arbitration.
Mr. Mills also engages in economic research and consulting
outside the context of litigation. He has assessed the anticipated
competitive effects of mergers and joint ventures on behalf of
government regulatory agencies and merging parties; developed
forecasts and strategic recommendations for government agencies
and clients involved with real estate development; and assisted
clients with the valuation of intangible assets and entire businesses.
Mr. Mills received a Bachelor of Science degree in economics and
history from Portland State University and a Master of Arts degree
in economics from the University of California at Santa Barbara.
-21-
About the Authors
DUBRAVKA K. TOSIC is an Economist and Principal at ERS
Group, a national leader in economic and statistical consulting.
She rejoined ERS Group in Spring 2010, after 12 years as a
Director in the Dispute Analysis practice of
PricewaterhouseCoopers, LLP in New York. Dr. Tosic has a
wealth of experience in leading and managing projects for private
and public sector clients, and their in-house and outside counsel,
involving economic and quantitative analyses and damage
calculations in a wide variety of complex disputes, litigation and
arbitration matters, and pro-active risk management and
compliance reviews.
Dr. Tosic’s primary areas of expertise are labor and employment
(employment discrimination litigation involving various employer
actions (e.g., hiring, promotion, termination), compensation
studies, reductions-in-force analyses, and wage and hour litigation)
and complex commercial litigation and disputes. She has provided
assistance as consulting expert or testifying expert to address
issues of class certification, liability, and estimated damages.
Additionally, Dr. Tosic has performed pro-active risk management
and compliance reviews, and management consulting projects
involving high-profile reform and transformational initiatives
involving process reviews and data analytics.
Dr. Tosic received her Ph.D. in economics from Florida State
University and her Bachelor’s degree from University of
Maryland, and previously worked at ERS Group from 1991-1996.
-22-
About ERS Group
ERS Group is the preeminent economic and statistical consulting
firm for analyses related to employment matters. Founded in 1981,
with offices in Tallahassee, Washington, D.C., San Francisco, and
Los Angeles, its statistically sound studies provide clients with a
better understanding of their organizations and decision-making
processes. Its research has been used by clients in high stakes
employment litigation and regulatory matters involving allegations
of discrimination in hiring, promotion and compensation. ERS
Group’s national reputation is founded on the unparalleled
experience of its economists and testifying experts. Its reach
extends to more than 3,000 clients, including Fortune 500
companies, prominent law firms, universities, government
agencies, and industry trade associations. Its experts also have
been asked to share their experience and knowledge with
regulatory agencies such as the Office of Federal Contract
Compliance and the Equal Employment Opportunity Commission.
About Micronomics
Micronomics is an economic research and consulting firm located
in Los Angeles, California. Founded in 1988, it is engaged in the
application of price theory, analysis of issues relating to resource
allocation, and assessment of real-world problems requiring
practical and sound solutions. Micronomics focuses on industrial
organization, antitrust, intellectual property, the calculation of
economic damages, employment issues, and the collection,
tabulation, and analysis of economic, financial and statistical data.
Clients include law firms, publicly and privately held businesses,
and government agencies. In January 2011, Micronomics joined
ERS Group.